Every model answers inside a fixed span of text called its context window — the instructions it was given, whatever was retrieved for it, the conversation so far, and your question, all held at once. Context engineering is the work of deciding what occupies that space. Not how politely the question is phrased, but which documents are fetched, which are discarded, how much of the history survives, and what sits closest to the question. The window is a desk with finite room, and someone has to choose what gets laid on it.
The most common failure is generosity. Faced with a limited window, the instinct is to fill it: retrieve everything that might be relevant, replay the whole conversation, include the entire page rather than the passage that answers. This degrades answers in a way that is genuinely hard to see, because nothing is missing. A fact sitting in the middle of a long context is drawn on less reliably than the same fact near the beginning or the end — an effect documented well enough to have earned a name. Add contradictory sources, or a stale document beside a current one, and the model has no way to know which to believe.
The discipline runs the other way: retrieve less and choose better. Fetch the three passages that answer the question rather than the fourteen pages that mention it. Summarise the conversation instead of replaying it. Put the material that matters where attention holds, near the question rather than buried at the midpoint. Keep every retrieved passage attributable, so the model can cite it rather than blend it into everything else. Each of these is a decision about what the model may look at, and each is made before the model runs.
Strip the machinery away and this is ordinary briefing. Hand a competent colleague forty pages and a vague question and you will get a hedged answer. Hand the same colleague the one paragraph that matters, and ask precisely, and you will get a useful one. Their intelligence did not change between the two attempts. What changed was what they were holding when you asked.
For anyone who owns a website rather than builds these systems, the half that matters is what gets retrieved in the first place. When an assistant answers a question about your business, your pages are the material somebody's retrieval step is choosing among. Whether a passage of yours survives that cut is decided by things you do control: whether it can be read at all, whether it carries a date, and whether it answers the question by itself instead of leaning on the paragraph above it. You cannot engineer someone else's context. You can decide what you are offering it.
